Speech recognition apparatus and method recognizing a speech from sound signals collected from outside

Abstract

A voice recognition system (10) for improving the toughness of voice recognition for a voice input for which a deteriorated feature amount cannot be completely identified. The system comprises at least two sound detecting means (16a, 16b) for detecting a sound signal, a sound source localizing unit (21) for determining the direction of a sound source based on the sound signal, a sound source separating unit (23) for separating a sound by the sound source from the sound signal based on the sound source direction, a mask producing unit (25) for producing a mask value according to the reliability of the separation results, a feature extracting unit (27) for extracting the feature amount of the sound signal, and a voice recognizing unit (29) for applying the mask to the feature amount to recognize a voice from the sound signal.

Description

TECHNICAL FIELD[0001]The present invention relates to a speech recognition apparatus and in particular it relates to a speech recognition apparatus that is robust to such speech that tends to deteriorate due to noises, input device specifications and so on.BACKGROUND OF THE INVENTION[0002]In general, a speech recognition apparatus in a real environment receives speech that deteriorates as it is mixed with noise and sound reverberations. The speech may also deteriorate depending on the specification of an input device. In order to cope with this problem, some approaches have been proposed for improving robustness of speech recognition by using such techniques as spectral subtraction, blind source separation and so on. One of such approaches proposed by M. Cooke et al. of Sheffield University is a missing feature theory (“Robust automatic speech recognition with missing and unreliable acoustic data”, SPEECH COMMUNICATION 34, p. 267-285, 2001 by Martin Cooke et al.). This approach aims...

AI Technical Summary

Benefits of technology

[0006]According to the invention, robustness of speech recognition can be improved because the value of the mask is generated according to the reliability of the result of separation of the speech from the sound signal by sound source.

[0009]According to a further aspect of the present invention, when there are multiple sound sources, the mask generation unit generates the value of the mask by increasing the reliability of separation result if (the signal is) closer to only one of the multiple sound sources.

Problems solved by technology

In the missing feature theory, there is often a difficulty in identifying deteriorated features when improvement of robustness of speech recognition is intended.

Images